A surveillance system that alerts every time it sees a person is worse than useless — it trains you to ignore it. Detection is the easy 10%. The hard, valuable 90% is deciding what actually deserves your attention. That's the layer I'm proudest of in Netra.
Detection is table stakes
Running YOLOv8 to find a person in a frame is a solved problem — plug it in and it works. But "a human is present" is almost never the thing you care about. A hallway sees dozens of people a day, all of them fine. If your system fires an alert on each one, you've built a very expensive way to generate noise, and the operator stops looking within a week.
The interesting question isn't is someone there? It's is this normal?
Scoring behavior, not presence
Netra's Behavioral Anomaly Signature Engine (BASE) goes past detection to judgment:
- Multi-factor threat scoring. Instead of a binary "person detected," it weighs multiple factors against configurable thresholds to produce a score — so "someone walked through at noon" and "someone is lingering somewhere they shouldn't be at 3 a.m." are treated as the different events they are.
- Time-weighted patrol heat-maps. Activity is aggregated over time and space, so the system builds a picture of where attention concentrates — normal patterns become the baseline that makes the abnormal stand out.
- Signal over volume. The whole design goal is fewer, better alerts. An alert that fires only when something is genuinely off is one an operator will actually trust.
The broader lesson
This is the shape of almost every real detection problem: the model that spots the thing is the beginning, not the end. The value is in the layer on top that turns raw detections into meaning — scoring, context, baselines, thresholds — so a human gets signal instead of a firehose.
Building BASE taught me that "the AI can detect X" is where the product starts. What you do with those detections — how you rank, contextualize, and filter them into something worth a person's attention — is where the engineering that matters lives.
The full anomaly-scoring breakdown and architecture are on the project page.
👉 See the system: www.divyakush.com/projects/netra
Divyakush Punjabi — Full-Stack & AI Systems Engineer
🌐 https://www.divyakush.com · 💼 LinkedIn · 💻 GitHub
Top comments (0)